JP2014115839A - Content recommendation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve user-friendliness when recommending contents.SOLUTION: A content recommendation device includes: a positive score calculation unit 31 which uses positive evaluation values of individual users for a plurality of contents to calculate positive scores being estimated probability values of positive evaluation values for contents which a target user has not evaluated, out of the plurality of contents; a negative score calculation unit 32 which uses negative evaluation values of individual users for a plurality of contents to calculate negative scores being estimated probability values of negative evaluation values for contents which the target user has not evaluated, out of the plurality of contents; and a recommendation content determination unit which determines contents to be recommended to the target user, on the basis of the positive scores calculated by the positive score calculation unit 31 and the negative scores calculated by the negative score calculation unit 32.

Description

  The present invention relates to a content recommendation device and a program.

  Conventionally, TV programs that match a person's interests and information (objective evaluation values) for identifying products that are likely to be purchased by a person have been collected from the person and a large number of unspecified persons. There is a technique for calculating based on the evaluation value data set (reference evaluation value) for the content set.

  As a technique for predicting the evaluation value of the target person from the evaluation value data for the content set collected for an unspecified number of people, refer to the evaluation values collected for an unspecified number of people and have similar preferences A method (collaborative filtering) for predicting the objective evaluation value by referring to and integrating the data is widely known, and its practical application is progressing. However, when a lot of missing values are included in the given reference evaluation value data, that is, when the reference evaluation value table is sparse, the content that can predict the target evaluation value is limited. It becomes a big problem. This problem is generally called the “sparse problem”.

  As a conventional technique capable of dealing with the sparse problem, for example, there is a related content search method using graph mining (see, for example, Non-Patent Document 1). This technique is a technique devised as a technique for predicting music that matches the taste of a specific user based on data collected by a user participation type music recommendation service. In this method, the evaluation value for each user's music provided by the service and tag information (genre) for each music are described in one graph format, and the transition relationship between the user and the music is expressed as follows: By using the RWR method (Random Walk with Restart), which is one of the graph mining methods, the content that matches the target user's preference is predicted.

Ioannis Konstas et al., “On Social Networks and Collaborative Recommendation,” Proc. 32nd ACM SIGIR 2009, pp. 195-202, 2009

  The RWR method according to the related art described above employs a method of describing an evaluation value as a “weight” at an edge of a graph to be analyzed. For this reason, when the evaluation value can take both a positive value and a negative value, one of the values (for example, the negative evaluation value) is ignored, and the other evaluation value (for example, the negative value) Only positive evaluation value). Therefore, the RWR method according to the prior art cannot predict the target evaluation value with high accuracy, and therefore cannot accurately estimate the content that matches the target user's preference. As a result, the user may not be able to obtain content recommendations that match the user's preference, and there is a problem that user convenience is low.

  One aspect of the present invention has been made in view of such circumstances, and an object thereof is to provide a content recommendation device and a program that can improve the convenience of the user at the time of content recommendation. is there.

  In order to solve the above-described problem, according to one aspect of the present invention, a positive evaluation value for a content that is not evaluated by a target user among the plurality of contents is obtained using a positive evaluation value of each user for each of the plurality of contents. Of the plurality of contents, the target user is evaluating using a positive score calculation unit that calculates a positive score that is a prediction probability value of the evaluation value and a negative evaluation value of each user for each of the plurality of contents. Based on a negative score calculation unit that calculates a negative score that is a predicted probability value of a negative evaluation value for no content, a positive score calculated by the positive score calculation unit, and a negative score calculated by the negative score calculation unit, Determination of recommended content for determining content recommended for the target user When a content recommendation apparatus comprising: a.

  According to an aspect of the present invention, in the content recommendation device, the recommended content determination unit may determine whether the evaluation value of the unevaluated content not evaluated by the target user is large or small based on the positive score and the negative score. A relationship is estimated, and content recommended to the target user is determined based on the estimated magnitude relationship.

  One aspect of the present invention is the content recommendation device, wherein the positive score calculation unit includes a similarity degree of preference between the users, a positive evaluation value for each content of each user, and a similarity degree between the contents. The negative score calculation unit refers to the preference similarity between the users, the negative evaluation value for each content of each user, and the similarity between the contents. Then, the negative score is calculated.

  One aspect of the present invention is the above content recommendation device, in which an evaluation value indicating an evaluation of each user for each of the plurality of contents is divided into the set of positive evaluation values and the set of negative evaluation values Section, each content and each user as a node, and a positive graph describing the set of positive evaluation values as edge weights connecting each node, and each content and each user as a node and the negative evaluation value And a creation unit that creates a negative graph describing the set of nodes as edge weights connecting each node, and the positive score calculation unit calculates the positive score based on the positive graph, and the negative score The score calculation unit calculates the negative score based on the negative graph.

  In order to solve the above-described problem, according to one embodiment of the present invention, a target user evaluates a plurality of contents using a positive evaluation value of each user for each of the plurality of contents. A positive score calculation step of calculating a positive score that is a predicted probability value of a positive evaluation value for a non-content, and a negative evaluation value of each user for each of the plurality of contents, and the purpose among the plurality of contents A negative score calculating step for calculating a negative score that is a predicted probability value of a negative evaluation value for content not evaluated by the user, a positive score calculated in the positive score calculating step, and a negative score calculating step Based on the negative score, the target user Is a program for executing the the recommended content determining step of determining a content to be recommended to.

  According to the present invention, it is possible to improve the convenience for the user at the time of content recommendation.

It is a block diagram which shows the structure of the content recommendation apparatus by one Embodiment of this invention. It is a flowchart for demonstrating operation | movement of the content recommendation apparatus by this embodiment. It is a flowchart for demonstrating operation | movement of the content recommendation apparatus by this embodiment. It is a conceptual diagram which shows an example of the basic evaluation value UI _base by this embodiment. It is a conceptual diagram which shows an example of the _target user evaluation value ui _target by this embodiment. Is a conceptual diagram showing an example of the reference evaluation value UI _ref according to the present embodiment. It is a conceptual diagram which shows an example of the positive graph _Gpos by this embodiment. It is a conceptual diagram which shows an example of the negative graph G _neg by this embodiment. It is a conceptual diagram which shows an example of the adjacency matrix A _pos by this embodiment. It is a conceptual diagram which shows an example of the adjacency matrix A _neg by this embodiment. It is a conceptual diagram which shows an example of the partial matrix II by this embodiment. It is a conceptual diagram which shows an example of the initial vector p ⁽⁰⁾ by this embodiment. It is a conceptual diagram which shows an example (in the case of (alpha) = 0.7, (epsilon) = ^10-10 ) of the update result _ppos finally obtained by this embodiment. An example of a finally obtained update result _{p neg} according to this embodiment is a conceptual diagram showing an (alpha = 0.7, the case of ε ^{= 10 -10).} It is a conceptual diagram which shows an example of the positive score s _pos by this embodiment. It is a key map showing an example of negative score s _neg by this embodiment. It is a conceptual diagram which shows an example of the integrated score s _total (when w _pos = w _neg = 1.0) by this embodiment. It is a conceptual diagram which shows an example (when threshold value (theta) = 1) of the recommended content list | wrist by this embodiment. It is the graph which compared the prediction accuracy about the top 1 of a recommendation program with the processing result by this embodiment, and the processing result using the conventional RWR method. It is the graph which compared the prediction accuracy about the top 5 of a recommendation program with the processing result by this embodiment, and the processing result using the conventional RWR method. It is the graph which compared the prediction accuracy about the top 10 of a recommendation program with the processing result by this embodiment, and the processing result using the conventional RWR method. It is the graph which compared the prediction accuracy about the top 20 of a recommendation program with the processing result by this embodiment, and the processing result using the conventional RWR method.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First, an outline of the present embodiment will be described. The content recommendation device 1 according to the present embodiment includes a data set of evaluation values of each unspecified person for a certain content set, and a data set of evaluation values of a specific person for some contents included in the previous content set. Evaluation value (hereinafter referred to as “objective evaluation value”) for a content for which a specific person's evaluation has not yet been given in the above-mentioned content set under the condition given with a set of (hereinafter referred to as “reference evaluation value”). ").

  In the content recommendation device 1 according to the present embodiment, the reference evaluation value is positive (for example, the degree of being a product to be purchased or a program to be viewed) and negative evaluation (for example, a product that is not desired to be purchased or a program that is not desired to be viewed). In the case where a scale along two opposite directions is obtained, the evaluation for each direction is evaluated and analyzed in an integrated manner. Thereby, the content recommendation device 1 can estimate the objective evaluation value with high accuracy. In particular, the content recommendation device 1 is robust even under conditions where there are many unevaluated data, that is, missing values (hereinafter referred to as “sparse conditions”) in the reference evaluation value that is the basis for calculating the objective evaluation value. Prediction is possible.

  In the situation where it is difficult to comprehensively investigate the evaluation values of an unspecified number of persons with respect to the same and a large amount of content sets, the content recommendation device 1 in the present embodiment responds positively to the evaluation values for each content. (Hereinafter referred to as “positive evaluation”) as well as anti-favorable reactions (hereinafter referred to as “negative evaluation”) are investigated and analyzed. Thereby, the content recommendation device 1 can predict the target evaluation value with higher accuracy.

  Here, for example, assuming that a broadcast program is the content to be analyzed, in order to predict a program that a specific person wants to watch, all household evaluations of all broadcast programs broadcast on terrestrial waves in Japan Exhaustive investigation of values is not practical. In many cases, the evaluation value for a part of broadcast programs investigated for the number of collaborators who can obtain a statistically superior number of responses is often used as basic data (ie, reference evaluation value).

  When the total number of contents to be analyzed or the total number of persons to be investigated is relatively small, it is possible to comprehensively investigate evaluation values (reference evaluation values) for all prepared contents. However, if one of the surveys is a large survey of more than a thousand, it is very difficult to accurately collect the evaluation values for all the surveyed content, That is, it is necessary to consider that missing values are included.

  Conventionally, the RWR method is known as a method capable of predicting a target evaluation value with high accuracy even in a situation where a lot of missing values are included in a given reference evaluation value. However, in the RWR method, the evaluation value for each content is described and analyzed as the “weight” of the graph. For this reason, favorable evaluation (positive evaluation) for each content can be described, but it is impossible in principle to describe anti-favorable evaluation (negative evaluation) at the same time, In order to integrate and analyze the information of both, a new device is required separately.

  On the other hand, the content recommendation device 1 according to the present embodiment describes both evaluation values in separate graphs when the reference evaluation value is given two directions of positive evaluation and negative evaluation. Then, the content recommendation device 1 integrates the two analysis results into one probability distribution after analyzing each graph independently. As a result, the content recommendation device 1 can predict a highly accurate objective evaluation value based on both positive and negative reference evaluation values.

  FIG. 1 is a block diagram showing a configuration of a content recommendation device 1 according to an embodiment of the present invention. In FIG. 1, the content recommendation device 1 includes a basic evaluation value recording unit 10, a graph creation unit 20, a purpose evaluation value prediction unit 30, and a recommended content determination unit 40.

  The basic evaluation value recording unit 10 records a basic content list, a reference user ID, and a reference user evaluation value. The basic content list is a list in which all contents to be analyzed are described. The reference user ID is a number for identifying each of an unspecified number of users who investigate an evaluation value for each content described in the basic content list. The reference user evaluation value is an evaluation value for each content described in the basic content list investigated for each reference user.

  The graph creation unit 20 receives the reference user evaluation values (hereinafter referred to as a list of evaluation values for each reference user) that are recorded for each reference user ID from the basic evaluation value recording unit 10 and the target user. The ID and the target user evaluation value are input, and a positive graph and a negative graph are created and output. The target user ID is identification information for identifying a target user (hereinafter referred to as a target user) to present a recommendation list. The target user evaluation value is a list of evaluation values made by the target user for each content in a subset of the previous content set.

  More specifically, the graph creating unit 20 includes a dividing unit 21 and a creating unit 22. The dividing unit 21 separates the reference user evaluation value into a positive evaluation value (a positive component of the reference evaluation value) and a negative evaluation value (a negative component of the reference evaluation value). The creation unit 22 uses each content and each user as a node, and connects each node with a positive evaluation value (a positive component of the reference evaluation value) of each user (reference user and target user) for each content. Creates a positive graph described as edge weights and a negative graph describing negative evaluation values (negative components of reference evaluation values) of each user (reference user and target user) for each content as edge weights And output.

  The objective evaluation value prediction unit 30 includes a positive score calculation unit 31 (positive score calculation unit) and a negative score calculation unit (negative score calculation unit) 32. The positive score calculation unit 31 describes the positive graph received from the graph creation unit 20 in the RWR execution parameters (specifically, the restart probability α and the convergence determination threshold value ε) separately given as input. Analysis is performed according to various set values, and a positive score is calculated and output as an analysis result. In addition, the negative score calculation unit 32 describes the negative graph received from the graph creation unit 20 in the RWR execution parameters (specifically, the restart probability α and the convergence determination threshold value ε) separately given as input. Analysis is performed according to the various set values, and a negative score is calculated and output as an analysis result.

  The positive score is obtained by approximating the predicted probability distribution of the positive evaluation value for the unevaluated content among the contents described in the basic content list of the target user. The negative score approximates the predicted probability distribution of negative evaluation values for the unevaluated content of the same user.

  The recommended content determination unit 40 includes an estimation unit 41 and a determination unit 42. The estimation unit 41 includes a recommended content determination parameter that is an aggregate of various setting values necessary for recommending specific content (programs and the like), a positive score previously calculated by the objective evaluation value prediction unit 30, and Based on the negative score, the target user calculates (estimates) the predicted probability value (magnitude relationship) for each content of the evaluation value (that is, the target evaluation value) for the content that has not been evaluated. The determination unit 42 rearranges the prediction probability values for each content in descending order, and outputs a list of content that gives a higher probability value, that is, a recommended content list.

Next, the operation of this embodiment will be described.
2 and 3 are flowcharts for explaining the operation of the content recommendation device 1 according to the present embodiment. The basic evaluation value recording unit 10 will be described as having an internal memory as an example.
First, the basic evaluation value recording unit 10 receives the basic content list and records it in the internal memory (step S100). Next, the basic evaluation value recording unit 10 receives the reference user ID and the reference user evaluation value, and records them in the internal memory (step S101). Then, the basic evaluation value recording unit 10 sends the basic evaluation value (table) UI _{base in} which the evaluation value is recorded for each reference user for each content described in the basic content list to the graph creating unit 20 (step S102).

FIG. 4 is a conceptual diagram illustrating an example of the basic evaluation value UI _base according to the present embodiment. In FIG. 4, in the basic evaluation value UI _base , evaluation values for the content contents # 1 to # 4 described in the basic content list are recorded for each reference user user # 1 and user # 2. Here, contents # 1 to # 4 are content IDs which are information for identifying the contents. In the illustrated example, evaluation values “0”, “+2”, “−1”, and “0” are recorded for the content contents # 1 to # 4 for the reference user user # 1. In addition, evaluation values “+2”, “−1”, “0”, and “+1” are recorded for the content users # 1 to # 4 for the reference user user # 2.

Returning to FIG. 2, the graph creating unit 20 next selects the target user ID for specifying the target user to be recommended and the evaluation value (purpose) for each content in the basic content list already evaluated by the target user. The user evaluation value list ui _target is received (step S103). Next, the graph creating unit 20 receives the basic evaluation value UI _base from the basic evaluation value recording unit 10 (step S104). Next, the graph creating unit 20 creates a reference evaluation value (table) UI _ref by combining the basic evaluation value (table) UI _base and the list of _target user evaluation values ui _target (step S105).

FIG. 5 is a conceptual diagram showing an example of the _target user evaluation value ui _target according to the present embodiment. In FIG. 5, an evaluation value that has already been evaluated for each content in the basic content list of the target user is recorded in the _target user evaluation value ui _target . In the illustrated example, evaluation values “+1”, “NA”, “−2”, and “NA” are recorded for the content users # 1 to # 4 for the target user # 0.

FIG. 6 is a conceptual diagram illustrating an example of the reference evaluation value UI _ref according to the present embodiment. In FIG. 6, the reference evaluation value UI _ref records the evaluation value for each content in the basic content list for each target user # 0, reference user user # 1, and user # 2. That is, the basic evaluation value UI _base shown in FIG. 4 and the _target user evaluation value ui _target shown in FIG. 5 are combined.

Next, the graph creating unit 20 separates the reference evaluation value UI _ref into a positive component and a negative component, and sets them as a positive graph G _pos and a negative graph G _neg (step S106). Specifically, the graph creating unit 20 extracts only one of the positive component (positive evaluation value) and the negative component (negative evaluation value) from each element of the created reference evaluation value UI _ref . Separate into two tables and fill each missing value NA with zero.

For each of the two obtained tables, the graph creation unit 20 uses each user and each content as a node, connects each user's node and the content node evaluated by the user with an edge, and sets the evaluation value to the relevant table. Assign a weight to an edge. However, for the negative component of the reference evaluation value UI _ref , the absolute value of the evaluation value (negative value) is given as the weight of the edge. Then, the graph creating unit 20 outputs the graph obtained for each of the positive component and the negative component of the reference evaluation value UI _ref as a positive graph G _pos and a negative graph G _neg .

FIG. 7 is a conceptual diagram illustrating an example of the positive graph G _pos according to the present embodiment. In FIG. 7, circles are nodes indicating each user and each content, and the thickness of an edge (line segment) connecting the nodes corresponds to the evaluation value. The numbers at each edge are evaluation values.

FIG. 8 is a conceptual diagram showing an example of the negative graph G _neg according to the present embodiment. In FIG. 8, as in FIG. 7, the circles are nodes indicating the users and the contents, and the thickness of the edge (line segment) connecting the nodes corresponds to the evaluation value. The numbers at each edge are evaluation values.

  Returning to FIG. 2, next, the graph creating unit 20 determines that the evaluation of the target user is not evaluated (content whose evaluation value is NA (missing value)) in the basic content list, and the target content to be predicted And the content ID indicating them is output as the target content ID (step S107).

Next, the objective evaluation value prediction unit 30 first reads the positive graph G _pos and the negative graph G _neg from the graph creation unit 20 (step S108). Next, the objective evaluation value prediction unit 30 reads each element of the RWR execution parameter, that is, the restart probability α and the convergence determination threshold value ε (step S109).

Next, the objective evaluation value prediction unit 30 obtains a matrix A ^* by substituting each element a _ij of the adjacency matrix A _pos extracted from the positive graph G _pos into the right side of the equation (2), and obtains the negative graph G each element _{a ij} of the adjacency matrix _{a neg} extracted from _neg, and substituted into the right side of equation (2) Find the matrix ^{a *} (step S110, S 111).

Specifically, the objective evaluation value prediction unit 30 follows the procedure described below for each of the positive graph G _pos and the negative graph G _neg, and determines the adjacency matrix A from the connection relationship between the nodes of the former graph. _The adjacency matrix A _neg is extracted from the connection relation of the latter graph, respectively.

FIG. 9 is a conceptual diagram illustrating an example of the adjacency matrix A _pos according to the present embodiment. As shown in FIG. 9, the adjacency matrix A _pos includes evaluation values between the users user # 0 to # 2 and the contents contents # 1 to # 4 extracted from the positive graph G _pos. It is a matrix.

FIG. 10 is a conceptual diagram illustrating an example of the adjacency matrix A _neg according to the present embodiment. As shown in FIG. 10, the adjacency matrix A _neg has an evaluation value between each user user # 0 to # 2 and each content content # 1 to # 4 extracted from the negative graph G _neg as an element. It is a matrix.

Then, these two matrices are substituted into the terms of the matrix A in the equation (1), respectively, and further, normalization processing is executed according to the equation (2), respectively, to obtain two types of normalized adjacency matrices A ^* . .

A matrix A represents an adjacency matrix determined by a connection relation between nodes of a given positive graph G _pos or negative graph G _neg and a connection strength (ie, edge weight), and four submatrices. It consists of UU, UI, IU, and II.

The submatrix UU is a matrix describing the preference similarity between the users, and the a row and b column components thereof are a vector v _{a in} which evaluation values for each content in the basic content list of the a-th user are arranged. , A correlation coefficient with the vector v _{b in} which evaluation values for the same content of the b-th user are arranged. However, for a set of users for which a negative correlation coefficient is obtained, the value of the corresponding matrix UU is replaced with 0.

Submatrix UI is a matrix representing the distribution of the evaluation value for each content for each user, that a row c columns component, c th element of the vector v _a (i.e., for the c-th content a th user Evaluation value). The submatrix IU is a transposed matrix in which the UI rows and columns are interchanged.

  The submatrix II is a matrix describing the degree of similarity between contents, and the c-row and d-column component is the degree of similarity between the c-th content and the d-th content. The similarity between the contents may be any index as long as it can express the magnitude relationship of the similarity based on the contents. For example, for broadcast program content, the ratio of the number of nouns that appear in common is calculated as the similarity for the pair of nouns that appear in the summary sentence of each program delivered by EPG (Electronic Program Guide). May be.

  FIG. 11 is a conceptual diagram showing an example of the submatrix II according to the present embodiment. In FIG. 11, in the partial matrix II, the degree of similarity with other contents is described for each of the contents contents # 1 to # 4. In the example shown in the figure, the match is “1”, and an index that is closer to “1” as it is similar and is smaller than “1” as it is not similar is used.

The matrix A ^* is a matrix obtained by normalizing the elements of the matrix A so that the sum of the elements for each column becomes 1 as shown in the equation (2).

Next, the objective evaluation value prediction unit 30 sets an initial vector in which, for each of the positive graph G _pos and the negative graph G _neg , only the value of the node corresponding to the target user ID is 1 and the other values are 0. p ⁽⁰⁾ is set (steps S111 and S115), and the adjacency matrices A _pos and A _neg are respectively normalized by using the normalized adjacency matrix A ^* and the initial vector p ⁽⁰⁾ , and the vector p according to equation (3). ^(T) is updated (steps S111 and S115).

FIG. 12 is a conceptual diagram illustrating an example of the initial vector p ⁽⁰⁾ according to the present embodiment. As shown in FIG. 12, the initial vector p ⁽⁰⁾ is a matrix in which only the value of the node corresponding to the target user ID is 1 and the other values are 0, as described above.

Specifically, for the two different matrices A ^* obtained in steps S110 and S114, the update process defined by Equation (3) is independently performed on the vector p representing the retention probability of each node.

Here, the vector p ^{(t + 1)} is a column vector that represents the expected value of the proximity to each user as viewed from the target user and a column vector that represents the expected value of the proximity to each content viewed from the target user. It represents the value of the vertically connected vector p in the t-th update step. The length corresponds to the total number of nodes described in a given positive graph G _pos or negative graph G _neg (that is, the sum of the total number of users handled by the system and the total number of contents).

The initial value p ⁽⁰⁾ of the vector is a vector (common fixed value) in which only the value of the element corresponding to the target user is set to 1 and all other elements are set to 0.

Next, the objective evaluation value prediction unit 30 determines whether or not the update result has converged, that is, the amount of change in the vector p ^(t) (specifically, p ^{(t + 1)} and the vector p ^{(t It} is determined whether or not the sum of the absolute values of the differences with respect to) is below the convergence determination threshold ε (steps S113 and S117). If the update result has not converged (NO in step S113, NO in S117), the process returns to steps S112 and S116, respectively, and the update process is repeatedly executed.

On the other hand, when the amount of change in p ^(t) falls below the convergence determination threshold value ε and the update is completed (YES in step S113, YES in S117), the objective evaluation value prediction unit 30 extracts from the positive graph G _pos. and executing the equation (1) to equation (3) for the adjacency matrix _{a pos,} finally obtained update result _{p pos,} as well as perform the same processing for the adjacency matrix _{a pos} extracted from negative graph _{G neg} Then, for each of the update results p _neg finally obtained, components corresponding to the target content ID are extracted in a vector format and output as positive score s _pos and negative score s _neg , respectively (step S118). ).

FIG. 13 is a conceptual diagram illustrating an example of the update result p _pos finally calculated by the objective evaluation value prediction unit 30 according to the present embodiment (when α = 0.7 and ε = 10 ⁻¹⁰ ). As illustrated in FIG. 13, the finally obtained update result p _pos corresponds to “users # 0 to # 2 and content contents # 1 to # 4”, “0.772, 0.065, 0. 072, 0.072, 0.011, 0.001, 0.008 ".

FIG. 14 is a conceptual diagram illustrating an example of the update result p _neg finally calculated by the objective evaluation value prediction unit 30 according to the present embodiment (when α = 0.7 and ε = 10 ⁻¹⁰ ). As illustrated in FIG. 14, the finally obtained update result p _neg corresponds to “0.772, 0.060, 0.00.0” corresponding to the users user # 0 to # 2 and the content contents # 1 to # 4. 053, 0.002, 0.007, 0.106, 0.000 ".

FIG. 15 is a conceptual diagram illustrating an example of a positive score s _pos according to the present embodiment. FIG. 15 illustrates a matrix including components corresponding to the target content IDs (= # 2, # 4) as an example of the positive score s _pos . For example, the positive score calculation unit 31 calculates the positive score s _pos shown in FIG. 15 and holds the calculated positive score s _pos .

FIG. 16 is a conceptual diagram illustrating an example of a negative score s _neg according to the present embodiment. FIG. 16 illustrates a matrix including components corresponding to the target content IDs (= # 2, # 4) as an example of the negative score s _neg . For example, the negative score calculation unit 32 calculates the negative score s _neg shown in FIG. 16 and holds the calculated negative score s _neg .

Next, the recommended content determination unit 40 reads the positive score s _pos and the negative score s _neg from the objective evaluation value prediction unit 30 (step S119), and sets the recommended content determination parameters (w _pos , w _neg , θ). Read (step S120). Next, the recommended content determination unit 40 approximates the positive score s _pos that approximates the target user's positive evaluation value for the prediction target content (target content) and the target user's negative evaluation value for the target content. Using the weighted difference with the negative score s _neg , the integrated score s _total is calculated according to the following equation (4) (step S121).

FIG. 17 is a conceptual diagram showing an example of the integrated score s _total (when w _pos = w _neg = 1.0) according to the present embodiment. 17, as an example of the total score _{s total,} object content ID (= # 2, # 4 ) matrix of component "0.004,0.008" corresponding to is shown. For example, the recommended content determination unit 40 calculates the integrated score s _total shown in FIG.

Then, the recommended content determination unit 40 sorts the values of the integrated score s _total in descending order, and outputs the IDs and scores of the content corresponding to the top θ items as the recommended content list (step S122).

FIG. 18 is a conceptual diagram showing an example of the recommended content list according to the present embodiment (when threshold θ = 1). In FIG. 18, as an example of the recommended content list, [user # 0. contents # 4, 0.008]. This is because 0.008 which is the component having the largest value among the components of the matrix representing the integrated score s _total in FIG. 17 and the corresponding target content ID # 4 are extracted as a recommended content list by the recommended content determination unit 40. Because it was done.

Next, the effectiveness of the above-described embodiment with respect to the prior art will be described.
19 to 22 are graphs showing a comparison between the processing result according to the present embodiment and the processing result using the conventional RWR method. FIG. 19 is a graph comparing the prediction accuracy for the top first recommended program between the processing result according to the present embodiment and the processing result using the conventional RWR method. FIG. 20 is a graph comparing the prediction accuracy of the top five recommended programs between the processing result according to the present embodiment and the processing result using the conventional RWR method. FIG. 21 is a graph comparing the prediction accuracy for the top 10 recommended programs between the processing result according to the present embodiment and the processing result using the conventional RWR method. FIG. 22 is a graph comparing the prediction accuracy for the top 20 recommended programs with the processing result according to the present embodiment and the processing result using the conventional RWR method.

  In order to obtain the graphs of FIGS. 19 to 22, five-stage evaluation of 1228 users for 300 programs prepared as contents (5: I want to see, 4: I want to see, 3: I can't say either, 2: I see too much) I do not want to see 1: I do not want to see). 19 to 22, the horizontal axis of the graph represents the data thinning rate (% Pruning rate). Here, the data thinning rate is the ratio of data to be hidden as a missing value (unrated program) among the program evaluation values referred to by the system. The vertical axis of the graph represents the prediction accuracy for the top k-th recommended program. Here, the prediction accuracy for the top k-th recommended program was the user evaluation of “4: Well want to see” among the top k programs (k is 1, 5, 10, or 20) of the calculated score. It is a ratio. In the graph, the prediction accuracy for the top k-th recommended program is expressed as% Precision @ k. Hereinafter, the prediction accuracy for the top k of the recommended programs is referred to as “P @ k”.

  In FIG. 19 to FIG. 22, black circles are data by a conventional standard RWR method. Prediction accuracy of program evaluation value by RWR method using only positive evaluation value (5 & 4) as adjacency matrix among five-level evaluation (average value of results of five trials using conventional standard RWR method) Is plotted as a black circle. White circles are data according to the RWR method of the present embodiment. As in the prior art, the positive evaluation value (5 & 4) and the negative evaluation value (2 & 1) of the five-level evaluation are reflected in different adjacency matrices, and the graph corresponding to each adjacency matrix is represented by an independent RWR algorithm. Prediction accuracy when integrating the results analyzed in (Randomly determine the number of programs corresponding to the set thinning rate out of the evaluation values for each user's 300 programs, and hidden the evaluation values as missing values Under the conditions, the average value of the results of five trials predicted by the method of the present invention is plotted as a white circle.

  19 to 22, the proportion corresponding to the line segment extending above the white circle and the black circle or the proportion extending below the line segment is a random sampling performed five times for each thinning rate setting. This is a ratio corresponding to the standard deviation for evaluating the variation in the prediction accuracy for the result.

  As shown in FIG. 19, in the case of the prediction accuracy (P @ 1) for the top first recommended program, there is no significant difference between the two. On the other hand, in the case of the prediction accuracy (P @ 5, P @ 10, P @ 20) including the second and subsequent recommended programs shown in FIG. 20, FIG. 21, and FIG. Also, the curve (white circle) of the result using the method according to the present invention is more than the fluctuation amount considering the vertical width of the error bar than the curve (black circle) of the result using the conventional standard RWR method. It can also be seen that it passes through the upper region. That is, it can be concluded that the method according to the present embodiment has a significantly higher prediction accuracy than the conventional method.

  As described above, in the content recommendation device 1 according to the present embodiment, the positive score calculation unit 31 refers to the preference similarity between users, the positive evaluation value for each content of each user, and the similarity between contents. The positive score is calculated by updating the expected value of proximity to each user from the target user and the expected value of proximity to each content viewed from the target user. Moreover, the negative score calculation unit 32 refers to the preference similarity between the users, the negative evaluation value for each content of each user, and the similarity between the contents, and from the target user to each user. A negative score is calculated by updating the expected value of proximity and the expected value of proximity to each content as viewed from the target user.

  The content recommendation device 1 according to the present embodiment conflicts with the reference evaluation value in the RWR method that is robust against the sparse problem even under the condition that the evaluation value is obtained only for a part of the content described in the basic content list. The process is performed independently for each of the components, that is, the positive evaluation value (positive graph) and the negative evaluation value (negative graph). Then, the content recommendation device 1 finally integrates each analysis result into one probability distribution. As described above, the content recommendation apparatus 1 performs the RWR method only on the positive evaluation value alone while being robust against the sparse problem by comprehensively analyzing the components of the conflicting evaluation for the same content. It is possible to predict the target evaluation value with higher accuracy than in the case of doing so. As a result, the content recommendation device 1 can recommend the content that matches the preference of the target user with high accuracy, so that the convenience of the target user can be improved.

Note that a system including a plurality of devices may process each process of the content recommendation device 1 of the present embodiment in a distributed manner by the plurality of devices.
Further, a program for executing each process of the content recommendation device 1 of the present embodiment is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed. The above-described various processes related to the content recommendation device 1 may be performed.

  Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” refers to a volatile memory (for example, DRAM (Dynamic) in a computer system serving as a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. Random Access Memory)) that holds a program for a certain period of time is also included. The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, a specific structure is not restricted to this embodiment. Each configuration in each embodiment, a combination thereof, and the like are examples, and the addition, omission, replacement, and other changes of the configuration can be made without departing from the spirit of the present invention. Further, the present invention is not limited by the embodiments, and is limited only by the scope of the claims.

DESCRIPTION OF SYMBOLS 1 Content recommendation apparatus 10 Basic evaluation value recording part 20 Graph preparation part 21 Dividing part 22 Creation part 30 Objective evaluation value prediction part 31 Positive score calculation part (positive score calculation part)
32 Negative score calculator (negative score calculator)
40 Recommended Content Determination Unit 41 Estimation Unit 42 Determination Unit

Claims (5)

Using the positive evaluation value of each user for each of the plurality of contents, a positive score calculation that calculates a positive score that is a prediction probability value of a positive evaluation value for the content not evaluated by the target user among the plurality of contents And
Negative that calculates a negative score that is a predicted probability value of a negative evaluation value for content that is not evaluated by the target user among the plurality of contents, using negative evaluation values of each user for each of the plurality of contents A score calculator;
Based on the positive score calculated by the positive score calculation unit and the negative score calculated by the negative score calculation unit, a recommended content determination unit that determines content recommended for the target user;
A content recommendation device comprising:   Based on the positive score and the negative score, the recommended content determination unit estimates a magnitude relationship between evaluation values of unevaluated content not evaluated by the target user, and based on the estimated magnitude relationship, The content recommendation device according to claim 1, wherein content recommended for the target user is determined. The positive score calculation unit calculates the positive score with reference to the similarity of preference between the users, the positive evaluation value for each content of each user and the similarity between the contents,
The negative score calculation unit calculates the negative score by referring to a preference similarity between the users, a negative evaluation value of each user for each content, and a similarity between the contents. The content recommendation device according to any one of claims 1 and 2. A dividing unit that divides an evaluation value indicating an evaluation of each user for each of the plurality of contents into the set of positive evaluation values and the set of negative evaluation values;
Each content and each user as a node, and the positive evaluation value set described as an edge weight connecting each node, and each content and each user as a node and the negative evaluation value set And a creation unit for creating a negative graph describing as an edge weight connecting each node,
The positive score calculation unit calculates the positive score based on the positive graph,
The content recommendation device according to any one of claims 1 to 3, wherein the negative score calculation unit calculates the negative score based on the negative graph. On the computer,
Using the positive evaluation value of each user for each of the plurality of contents, a positive score calculation that calculates a positive score that is a prediction probability value of a positive evaluation value for the content not evaluated by the target user among the plurality of contents Steps,
Negative that calculates a negative score that is a predicted probability value of a negative evaluation value for content that is not evaluated by the target user among the plurality of contents, using negative evaluation values of each user for each of the plurality of contents A score calculating step;
Based on the positive score calculated in the positive score calculation step and the negative score calculated in the negative score calculation step, a recommended content determination step for determining content recommended for the target user;
A program for running

Images